[Feature Request] Add multimodal/vision support for Claude and other vision-capable models

[NoxzRCW]

Problem

Currently, 9Router does not transmit images to upstream providers, even when using vision-capable models like Claude Sonnet 4.5.

Test Case

I tested sending an image via the OpenAI-compatible API format:

curl -X POST http://192.168.11.233:20128/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer sk-xxx" \
  -d '{
    "model": "kr/claude-sonnet-4.5",
    "messages": [
      {
        "role": "user",
        "content": [
          {"type": "text", "text": "What do you see in this image?"},
          {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}}
        ]
      }
    ]
  }'

Expected: Claude should describe the image
Actual: Claude responds "I don't see any image in your message"

Impact

This prevents using 9Router with:

• Claude Sonnet/Opus (vision support)
• GPT-4V/GPT-5 (vision support)
• Gemini (multimodal)
• Any other vision-capable models

Use Case

I'm using 9Router with OpenClaw (personal AI assistant) and need vision support for analyzing screenshots, diagrams, and images sent via Telegram/WhatsApp.

Proposed Solution

1. Detect image_url content blocks in OpenAI format
2. Translate to provider-specific format:
   • Claude: Convert to Anthropic's vision format with source.type: "base64"
   • OpenAI: Pass through as-is
   • Gemini: Convert to inline_data format
3. Preserve image data through the routing pipeline

Environment

• 9Router version: 0.2.99
• Model tested: kr/claude-sonnet-4.5 (via Kiro provider)
• Client: OpenClaw via custom provider config

Additional Context

The /v1/models endpoint doesn't expose model capabilities (text vs multimodal), which makes it hard for clients to know which models support images.

Would be great to also add:

• Model capability metadata in /v1/models response
• Documentation on multimodal support per provider

Thanks for this amazing project! 🚀

[attid]

claude work for me. OpenAI not work =(

[East-rayyy]

The root cause is in the OpenAI-to-Claude and OpenAI-to-Gemini request translators. Both only handle data: base64 image URLs — HTTP/HTTPS URLs are silently dropped with no error.

In openai-to-claude.js, the image_url handler matches /^data:([^;]+);base64,(.+)$/ and has no else branch for HTTP URLs. Same pattern in geminiHelper.js — it checks startsWith("data:") and skips everything else.

PR #344 adds HTTP/HTTPS URL support to both translators — Claude's source: { type: "url" } format and Gemini's fileData: { fileUri } format.